The Pure Dispatcher Pattern: Decoupling Orchestration from Computation in Distributed AI Workloads

Preprint — not peer reviewed. Distributed computing frameworks such as Ray, Spark, and Dask conventionally integrate orchestration and computation in the same execution layer. This work argues that, for embarrassingly parallel workloads on heterogeneous hardware accelerators (the EPHA class), this integration is the structural source of irreducible coordination overhead — overhead that is architectural in origin and cannot be removed by parameter tuning. We formalize the Pure Dispatcher Pattern (PDP): an architectural restriction on coordinator behavior in which the coordinator is stateless, data-agnostic, fault-transparent, and partition-exclusive — it dispatches only work-unit identifiers and never touches computation data, which workers retrieve and process locally. Under the stated coordinator model we prove (i) a coordination-overhead optimality result, (ii) a separation theorem showing that every data-touching coordinator is provably separated from PDP by a factor that grows with pipeline depth, intermediate size, and per-byte transfer cost, (iii) a minimal axiomatization of the four defining properties, and (iv) a makespan/speedup model that predicts end-to-end behavior from the measured coordination tax. We instantiate PDP in Ray and evaluate it on a production EPHA workload (OCR-to-RAG over 11,368 pages across 2,246 documents). A controlled comparison at matched resolution and rasterization attributes a 2.29× speedup to coordinator design and a 77.5× reduction in measured coordination overhead; total speedup with resolution optimization reaches 69.9×. A corpus-scaling experiment (500 to 11,368 pages) is consistent with the predicted overhead growth. PDP is not a new scheduler, a new OCR system, or a replacement for Ray/Spark/Dask; it is an architectural constraint on coordinator behavior and its formal consequences. The OCR-to-RAG pipeline is used only as a validation workload. Multi-GPU (Architecture 3) scaling is an analytical projection, not an empirical result.